[
https://issues.apache.org/jira/browse/SPARK-4452?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14216691#comment-14216691
]
Matei Zaharia commented on SPARK-4452:
--------------------------------------
BTW I've thought about this more and here's what I'd suggest: try a version
where each object is allowed to ramp up to a certain size (say 5 MB) before
being subject to the limit, and if that doesn't work, then maybe go for the
forced-spilling one. The reason is that as soon as N objects are active, the
ShuffleMemoryManager will not let any object ramp up to more than 1/N, so it
just has to fill up its current quota and stop. This means that scenarios with
very little free memory might only happen at the beginning (when tasks start
up). If we can make this work, then we avoid a lot of concurrency problems that
would happen with forced spilling.
Another improvement would be to make the Spillables request less than 2x their
current memory when they ramp up, e.g. 1.5x. They'd then make more requests but
it would lead to slower ramp-up and more of a chance for other threads to grab
memory. But I think this will have less impact than simply increasing that free
minimum amount.
> Shuffle data structures can starve others on the same thread for memory
> ------------------------------------------------------------------------
>
> Key: SPARK-4452
> URL: https://issues.apache.org/jira/browse/SPARK-4452
> Project: Spark
> Issue Type: Bug
> Affects Versions: 1.1.0
> Reporter: Tianshuo Deng
> Assignee: Tianshuo Deng
> Priority: Blocker
>
> When an Aggregator is used with ExternalSorter in a task, spark will create
> many small files and could cause too many files open error during merging.
> Currently, ShuffleMemoryManager does not work well when there are 2 spillable
> objects in a thread, which are ExternalSorter and ExternalAppendOnlyMap(used
> by Aggregator) in this case. Here is an example: Due to the usage of mapside
> aggregation, ExternalAppendOnlyMap is created first to read the RDD. It may
> ask as much memory as it can, which is totalMem/numberOfThreads. Then later
> on when ExternalSorter is created in the same thread, the
> ShuffleMemoryManager could refuse to allocate more memory to it, since the
> memory is already given to the previous requested
> object(ExternalAppendOnlyMap). That causes the ExternalSorter keeps spilling
> small files(due to the lack of memory)
> I'm currently working on a PR to address these two issues. It will include
> following changes:
> 1. The ShuffleMemoryManager should not only track the memory usage for each
> thread, but also the object who holds the memory
> 2. The ShuffleMemoryManager should be able to trigger the spilling of a
> spillable object. In this way, if a new object in a thread is requesting
> memory, the old occupant could be evicted/spilled. Previously the spillable
> objects trigger spilling by themselves. So one may not trigger spilling even
> if another object in the same thread needs more memory. After this change The
> ShuffleMemoryManager could trigger the spilling of an object if it needs to.
> 3. Make the iterator of ExternalAppendOnlyMap spillable. Previously
> ExternalAppendOnlyMap returns an destructive iterator and can not be spilled
> after the iterator is returned. This should be changed so that even after the
> iterator is returned, the ShuffleMemoryManager can still spill it.
> Currently, I have a working branch in progress:
> https://github.com/tsdeng/spark/tree/enhance_memory_manager. Already made
> change 3 and have a prototype of change 1 and 2 to evict spillable from
> memory manager, still in progress. I will send a PR when it's done.
> Any feedback or thoughts on this change is highly appreciated !
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional commands, e-mail: issues-h...@spark.apache.org
